Bracketing Paradoxes Revisited: Fresh Experimental Insights and a New Reanalysis

Stuttgart 14.01.2025

Anna Prysłopska

anna.pryslopska@ling.uni-stuttgart.de

Universität Stuttgart

Institut für Linguistik

Titus von der Malsburg

titus.von-der-malsburg@ling.uni-stuttgart.de

Universität Stuttgart

Institut für Linguistik

Nominal compounds

der Obstbaum
the fruit tree

das Obst
the fruit

der Baum
the tree

Singer-Songwriter, Denkfehler, Kuscheltier

The subject of the research I would like to present today is phrases which consist of a noun compound and an adjective. When I talk about nominal composites, I mean NN composites. I would like to exclude other composites (verb+noun, adjective+noun, etc.). singer-songwriter, fallacy, stuffed animal

Modification with adjectives

Same syntactic structure, seemingly different semantic structures, which violates the Principle of Compositionality.

the Ukrainian tomb of a (Roman) prince

the tomb of a Ukrainian prince

Bracketing paradox! Not compositional?

Adjectives that modify noun phrases adapt their moprhological moprhological properties to the head, which is always at the very end. I will refer to the head noun as the second noun and N2 in the course of the lecture.

Adjective and NN compound phrases have two readings. Normally, the adjective modifies the main noun (also called the iconic reading). “Celtic” modifies ‘grave’ and the nationality of the prince is unknown.

The second reading (anti-iconic) is rare, but does occur. Here the non-head is modified (I call it N1, even though this is not necessarily just a noun) In this “Celtic” modifies “prince” and it is a grave of a Celtic prince.

Both constructions have the synactic parenthesis (on the left), but the one in the second reading the semantic bracketing seems to differ from the syntactic bracketing. Does this violate the principle of compositionality?

Despite the fact that these constructions are “non-compositional” and are also described as gross nonsense, they occur very frequently in both written and spoken language.

The canonical reading is the norm. Some phrases are ambiguous or are even preferentially interpreted as a bracketing paradox.

N2

N1

steiniger Fahrradweg stony bike path
→ canonical reading

königliche Hochzeitsfeier royal wedding celebration
→ canonical and bracketing paradox reading

saurer Kirschbaum sour cherry tree
→ only bracketing paradox reading

?? dreiköpfiger Familienvater 3 headed family father
→ father of a family of 3 vs. Cerberus

Some phrases clearly have an N2/iconic/canonical reading, while others have a clear N1/anti-iconic/BP reading. Others are somewhere in between. There is also no regularity to how natural or acceptable the parenthesis paradoxes are.

What licenses the bracketing paradox reading?

How are bracketing paradoxes processed?

There is hardly any psycholinguistic research on this topic!

We asked ourselves: How do we model these phrases and how are they processed?

Mistakes, humor, pragmatics,

plausibility, context, world knowledge,

semantic transparency, language economy,

information density, adjective semantics,

lexicalization of the compound noun,

internal compound semantics,

semantic agreement,

morphosyntactic congruency

between adjective and noun

Experiment 1

Do the semantic agreement and the morphosyntactic congruence between the adjective and the individual nouns affect the acceptability of the entire phrase?

In the first experiment, we investigated whether the acceptability of the entire adjective-NN compound phrase is influenced by the semantic and the morphosyntactic match between the adjective and the individual nouns

Method

We collected phrases from linguistic literature and newspaper articles

204 adjective+NN compound phrases from literature, newspaper articles

Interpretation: canonical, ambiguous, bracketing paradox

Constructions where the adjective does not match any of the nouns → unlikely to be used

saurer Kirsch
sour cherry

AN1

saurer Baum
sour tree

AN2

saurer Kirschbaum
sour cherry tree

AN1N2

3 conditions

Schmidt (1993) hat ähnliche Phänomene empirisch untersucht und dabei einige der Beispiele von Bergmann verwendet

3 lists, 36 participants on Prolific, 3 scales

Scales

Comprehensibility judgments were highest, but the scales were highly correlated (R ≥ 0.94)
→ measure related concepts

[0-1] scaled mean for analyses

Semantic agreement

Each point is a phrase

Semantic agreement: Predictions

Strict compositionality: N2 is the head noun.

AN1N2 ratings rise when AN2 ratings rise
high agreement → canonical reading

AN1 ratings should not influence AN1N2 ratings
high agreement → bracketing paradox potential

Semantic agreement: Results

Strict compositionality: N2 is the head noun.

AN1N2 ratings rise when AN2 ratings rise
high agreement → canonical reading

AN1N2 ratings rise when AN2 ratings rise
high agreement → canonical reading

AN1 ratings should not influence AN1N2 ratings
high agreement → bracketing paradox potential

AN1N2 ratings rise a bit when AN1 ratings rise
high agreement → bracketing paradox potential

AN1 × AN2 interaction
High AN1 and AN2 agreement reduced AN1N2 ratings.

Morphosyn. congruency: Predictions

If the adjective agrees with N1 and N2, this could increase the likelihood that readers will accept a BP interpretation. Supported by morphosyntax. We have annotated the agreement between the adjective and N1. schweres Unwetter das schwere Unwetter schweres Unwetter

The morphosyntactic features of the adjective must agree with N2.

The morphosyntactic congruency between the adjective and N1 increases bracketing paradox potential.

schwerefem/neut/plsevere	Unwetterneut/plstorm	warnungfemwarning	✔
königlichefem/neut/plroyal	Hochzeitsfemwedding	feierfemcelebration	✔
saurermascsour	Kirschfemcherry	baummasc tree	✘
dreiköpfigermasc 3-head	Familienfem/plfamily	vatermascfather	✘

Morphosyn. congruency: Results

The morphosyntactic features of the adjective must agree with N2.

The morphosyntactic congruency between the adjective and N1 had no effect.

Experiment 1 summary

When adjective does not match the head noun semantically, the first noun becomes an attractive target for the adjective.

Both nouns have a positive effect on the acceptability of an adjective and NN-compound phrase.

If the adjective matches both nouns, there is a conflict or competition between them.

The morphosyntactic congruence between the adjective and the non-head noun has no effect.

Comprehensibility, naturalness, and stylistic form are related concepts.

neidischer Zickenkrieg 🗲 jealous catfight

semantic agreement ✔ morphosyntactic congruency ✘

Experiment 2

Which noun in a compound is the adjective attached to?

Next, we asked ourselves which noun does the adjective ACTUALLY attach to

answer options: n1, n2, unclear (may opt out of interpreting altogether)

235 phrases in 1 condition, 20 participants on Prolific

Predictions

Strict compositionality: The adjective attaches to N2.

N2 is chosen more often than N1

N1 is chosen more often than N2

N1 is chosen as often as N2

? chosen often

Adjective N1 morphosyntactically

Canonical

Bracketing paradox

Ambiguous

no interpretation

N1 attachment facilitated

The semantic agreement ratings from Experiment 1 should align with the selected attachment site.

Results: Attachment preference

Our corpus consisted of canonical and BP examples, so not unexpected to find N2 attachments

N2 was chosen overall more often than N1

N1 was sometimes chosen more often than N2

N1 was sometimes chosen as often as N2

? chosen often

grammatical, canonical

bracketing paradox, not compositional?

ambiguous or bracketing paradox?

almost always interpreted

Attachment and AN2 ratings: Predictions

The compatibility ratings should match the attachment site.
A N2 → ¬N1

The compatibility ratings should match the attachment site.
A N1 → N1

Attachment and AN2 ratings: Results

Semantic agreement between adjective to N2 reduces N1 attachment.

Semantic agreement between adjective to N1 increases N1 attachment.

The compatibility ratings should match the attachment site.
A N2 → ¬N1

The compatibility ratings match the attachment site.
A N1 → ¬N1

The compatibility ratings should match the attachment site.
A N1 → N1

The compatibility ratings match the attachment site.
A N1 → N1

Morphosyntactic congruency: Predictions

The morphosyntactic match between adjective and N1 should facilitate a bracketing paradox reading.

Morphosyntactic congruency: Results

The morphosyntactic match between adjective and N1 facilitates a bracketing paradox reading, but its influence is dwarfed by semantic agreement.

Experiment 2 summary

The results are consistent with Experiment 1.

Participants did not opt out of interpreting the phrases, but often disagreed what the interpretation is.

Most phrases received a canonical interpretation, but N1 was often an appropriate attachment noun.

Semantic agreement determines interpretational preferences, but morphosyntactic adjective and N1 match also facilitates a bracketing paradox reading.

semantic agreement ✔ morphosyntactic congruency ✔

Processing

Reaction times and ratings

Participants needed a little more time to decide in favor of N1.

Semantic agreement between adjective and noun shortened the reaction time.

If none of the nouns matched the adjective, participants took a long time to decide.

In Experiment 2, semantic agreement between adjective and noun accelerated the decision making.

In Experiment 1, there was a conflict or competition between the nouns.

\(\rightarrow\) ambiguity resolution, reinterpretation

Garden-path effect?

Experiment 4

Does context improve the interpretation of bracketing paradoxes?

We felt the phrases sounded better when embedded in a sentence or context.

Im Elternabend berichtete der dreiköpfige Familienvater von seinen Erfahrungen im Zwiespalt zwischen Familie und Beruf.

At the parents' evening, the father of three reported on his experiences of the conflict between family and career.

Ratings: Predictions

Bracketing paradoxes are embedded in a sentence will be rated higher.

Morphosyntactic congruency ??

Ratings: Results

Bracketing paradoxes are embedded in a sentence will be rated higher.

Bracketing paradoxes are embedded in a sentence are rated somewhat higher; canonical phrases also benefit.

Bracketing paradoxes received lower ratings.

Morphosyntactic congruency ??

Morphosyntactic congruency ??

Scales: Predictions

Experiment 1

Scales: Results

Experiment 1

Experiment 4

Experiment 4 summary

Sentence context improves phrase acceptability overall, but bracketing paradoxes benefit more than canonical phrases.

Comprehensibility, naturalness, and stylistic form are related concepts.

The findings are consistent with Experiments 1 and 2.

semantic agreement morphosyntactic congruency ✘ communicative relevance ✔

Experiment 5

Does context influence which noun in a compound is the adjective attached to?

235 phrases, 1 condition, 2 lists, 34 participants on Prolific

Attachment preference: Predictions

Same as in experiment 2

Strict compositionality: N2 is the head noun.

Experiment 2 → There will be a variety of interpretations.

N2 is chosen more often than N1

N1 is chosen more often than N2

N1 is chosen as often as N2

? chosen often

Canonical

Bracketing paradox

Ambiguous

no interpretation

Sentence context licenses a bracketing paradox interpretation.

Attachment preference: Results

Phrase ratings and attachments: Predictions

Experiment 2 The compatibility ratings should match the attachment site.
A N2 → ¬N1

Experiment 2 The compatibility ratings should match the attachment site.
A N1 → N1

Phrase ratings and attachments: Results

Experiment 2 The compatibility ratings should match the attachment site.
A N2 → ¬N1

Experiment 5 The compatibility ratings match the attachment site.
A N2 → ¬N1

Experiment 2 The compatibility ratings should match the attachment site.
A N1 → N1

Experiment 5 The compatibility ratings match the attachment site.
A N1 → N1

AN1 × AN2 interaction N1 attachments were higher when A N1 and A 🗲 N2

Sentence ratings and attachments: Predictions

The attachments in sentences should correspond to attachments in bare phrases.
Experiment 5 ≈ Experiment 2

Sentence ratings and attachments: Results

The attachments in sentences should correspond to attachments in bare phrases.
Experiment 5 ≈ Experiment 2

The attachments in sentences corresponded to attachments in bare phrases.
Experiment 5 ≈ Experiment 2

When sentence ratings were high, more N1 attachments in bare phrases translated to more N1 attachments in sentences.

Conclusions

FIXME there seems to be no point to experiment 6

The bracketing paradox interpretation is determined primarily by semantic agreement.

Morphosyntactic agreement between the adjective and non-head noun plays a smaller role.

Bracketing paradoxes benefit disproportionately from sentence context compared to canonical phrases.

Processing bracketing paradoxes resembles ambiguity resolution in garden-path sentences.

Future work

Mistakes, humor, pragmatics,

plausibility, context, world knowledge,

semantic transparency, language economy,

information density, adjective semantics,

lexicalization of the compound noun,

internal compound semantics,

semantic agreement,

morphosyntactic congruency

between adjective and noun

Information density

Väter mit einem Kind und solche mit mehreren Kindern erhalten unterschiedliche staatliche Unterstützungen.
Fathers with one child and those with several children receive different levels of state support.

Männer mit drei Köpfen haben oft Schwierigkeiten, einen Job zu finden, um ihre Familie zu unterstützen.
Men with three heads often have difficulty finding a job to support their family.

Der dreiköpfige Familienvater bekommt deutlich weniger Geld.
The father of three receives significantly less money.

Adjective semantics

Olga is a beautiful dancer.

Olga dances beautifully.
5/10 looks
¬ intersective

Olga is beautiful.
5/10 dancing skills
intersective

Adjective semantics

gegrillter Gemüsemix
grilled vegetable mix
participial

tropischer Pflanzenliebhaber
tropical plant lover
qualitative

lateinische Wortkunde
Latin lexicography
relational

Compound noun interaction

A Catholic church leader is probably Catholic. Pastafarian

katholisches Kirchenoberhaupt

(Self-paced/Eyetracking during) reading study

Die Forscher suchten nach wissenschaftlichen Versuchspersonen für ihre Sprachstudien.

Die Forscher suchten nach wissenschaftlichen Versuchspersonen für ihre Sprachstudien.
The researchers were looking for participants for their language studies.

What measures? The phrases and sentences are different in length, how to control for the (sentence) context?

Thank you

Bonus: Experiment 3

Can the adjective modify both nouns in a compound noun?

Is it only ever either N1 or N2?

What if the adjective attaches to N1, N2, or both?

Ukrainian prince(ly) tomb

More answer options →
More freedom →
More comprehensive picture

answer options: n1, n2, unclear (may opt out of interpreting altogether)

235 phrases in 1 condition, 20 participants on Prolific

Attachment preference: Predictions

Same as in experiment 2

Strict compositionality: N2 is the head noun.

Experiment 2: There will be a variety of interpretations.

N2 is chosen more often than N1

N1 is chosen more often than N2

N1 is chosen as often as N2

? chosen often

Canonical

Bracketing paradox

Ambiguous

no interpretation

Attachment preference: Results

Our corpus consisted of canonical and BP examples, so not unexpected to find N2 attachments

participants that made use of 2 simultaneous yes/no answers

More ✔ answers than in Experiment 2 and attachment to both nouns.

As in Experiment 2, there were predominantly N2 answers, but N1 was chosen equally or more often for some items.

Yes-answer likelihood: Predictions

Head position, good semantic agreement between the adjective and a noun → ✔ answers

Good semantic agreement between the adjective and both nouns → ✔✔ answers

(Ambiguity, attachment to both nouns?)

✔-answer likelihood: Predictions

The semantic compatibility ratings should correspond to the reference noun.

Adjective N1 → fewer ✔ for N2

Adjective N2 → fewer ✔ for N1

✔-answer likelihood: Results

Semantic agreement ratings align with the attachment preference.

Good semantic agreement between the adjective and a noun → ✔ answers

Adjective N1 → fewer ✔ for N2

Adjective N2 → fewer ✔ for N1

AN1 must be a great fit.

AN2 must fit a bit.

✔✔-answer likelihood: Results

In 13% of cases, the participants gave both nouns ✔

katholisches Kirchenoberhaupt
Catholic church leader

keltisches Fürstengrab
Celtic prince(ly) tomb

militärischer Besatzungsstützpunkt
military occupation base

✔✔-answer likelihood: Results

Semantic adjective-N1 agreement increases double-✔

Semantic adjective-N2 agreement has a weaker influence

If both nouns fit the adjective, double-✔ probability is highest.

Französisches Revolutionstribunal
French Revolutionary Tribunal

umweltfreundliche Windenergie
environmentally friendly wind energy

blaues Glasauge
blue glass eye

Summary

The canonical N2 interpretation is the default one, as expected by compositionality rules.

Nouns that semantically match the adjective receive positive responses. The hurdle is very high for bracketing paradoxes and very low for canonical phrases.

Ony if the adjective and N1 fit well semantically is a positive response be given for both nouns.

Analyses

Experiment 1

Parameter	Est.	CrI low	CrI high	Rhat	ESS
Intercept	-3.78	-5.33	-2.34	1	3608.82
AN1 rating	3.06	1.36	4.88	1	3599.17
AN2 rating	5.99	4.28	7.82	1	3554.80
Morphosynt. match	-0.44	-1.50	0.63	1	4343.46
AN1 rating × AN2 rating	-3.60	-5.77	-1.54	1	3599.28
AN1 rating × Morphosynt. match	0.62	-0.29	1.52	1	5127.36
AN2 rating × Morphosynt. match	0.04	-0.89	0.98	1	5097.54

Experiment 2

Parameter	Est.	CrI low	CrI high	Rhat	ESS
Intercept	-1.87	-3.32	-0.44	1	4193.23
AN1 rating	3.84	2.23	5.46	1	4250.93
AN2 rating	-1.97	-3.53	-0.41	1	4273.96
Morphosynt. match	-0.24	-1.49	1.00	1	4819.33
AN1 rating × AN2 rating	-1.06	-2.84	0.76	1	4412.81
AN1 rating × Morphosynt. match	0.33	-0.70	1.40	1	5380.05
AN2 rating × Morphosynt. match	0.58	-0.51	1.67	1	5128.95

Experiment 3: One “yes” answer

Parameter	Est.	CrI low	CrI high	Rhat	ESS
Intercept	1.13	0.15	2.12	1	4834.79
AN1 rating	-2.81	-3.94	-1.72	1	4793.24
AN2 rating	2.95	1.85	4.02	1	4768.11
Noun position	-2.70	-3.42	-1.98	1	5056.62
AN1 rating × AN2 rating	0.47	-0.83	1.77	1	4811.83
AN1 rating × Noun position	5.86	5.25	6.47	1	5557.05
AN2 rating × Noun position	-5.06	-5.70	-4.42	1	5501.25

Experiment 3: two “yes” answers

Parameter	Est.	CrI low	CrI high	Rhat	ESS
Intercept	-3.11	-3.88	-2.38	1	1960
AN1 rating	1.05	0.70	1.45	1	3984
AN2 rating	0.40	0.07	0.78	1	3469
AN1 rating × AN2 rating	0.33	0.02	0.63	1	3956